Function llama_model_load

llama.cpp:4220–4271

}

static bool llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params, const llama_context_params * cparams) {
    try {
        llama_model_loader ml(fname, params.use_mmap);

        if (ml.sparse_deriv == GGML_SPARSE_INFERENCE) {
            LLAMA_LOG_INFO("%s: PowerInfer model loaded. Sparse inference will be used.\n", __func__);
        }

        model.hparams.vocab_only = params.vocab_only;
        model.sparse_deriv = ml.sparse_deriv;

        llm_load_arch   (ml, model);
        llm_load_hparams(ml, model);
        llm_load_vocab  (ml, model);

        llm_load_print_meta(ml, model);

        if (model.hparams.n_vocab != model.vocab.id_to_token.size()) {
            throw std::runtime_error("vocab size mismatch");
        }

        if (params.vocab_only) {
            LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
            return true;
        }

        if (llama_use_sparse_inference(&model)) {
            if (params.n_gpu_layers > 0) {
                LLAMA_LOG_WARN("%s: sparse inference ignores n_gpu_layers, you can use --vram-budget option instead\n", __func__);
                return false;
            }
#if defined GGML_USE_CUBLAS
            llama_set_vram_budget(params.vram_budget_gb, params.main_gpu);
#endif
            llm_load_sparse_model_tensors(
                ml, model, cparams, params.main_gpu, vram_budget_bytes, params.reset_gpu_index, params.disable_gpu_index,
                params.use_mlock, params.progress_callback, params.progress_callback_user_data
            );
        } else {
            llm_load_tensors(
                ml, model, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.use_mlock,
                params.progress_callback, params.progress_callback_user_data
            );
        }

    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("error loading model: %s\n", err.what());
        return false;
    }

    return true;
}

//
// llm_build